XGBoost-based Hurricane Wind Reconstruction

Abstract:

This study describes a new XGBoost-based wind reconstruction model that can improve upon hurricane wind risk assessments from existing parametric wind models. Hurricane wind risk is measured by the probability of storm winds exceeding given thresholds at a particular location. Such probabilities are commonly derived from parametric surface wind models that consist of a radially symmetric wind field to which a small left-to-right asymmetry is added due to storm motion. Observed surface wind fields, however, can be highly variable with differing asymmetries, especially so for storms that encounter strong vertical wind shear or that undergo extra-tropical transition. The discrepancy between parametric surface wind fields and observations raises a question of how missing wind asymmetries can be added to existing parametric wind models. We address this question by developing an XGBoost-based wind reconstruction model. The input variables of the XGBoost model are storm and environment features and a reference symmetric wind field that is computed by a parametric wind model. The output of the XGBoost model is the predicted difference between the reference symmetrical wind fields and the observations. The predicted difference is represented using a small number of Laplacian eigenfunctions as basis functions. We train and test the model using HWIND from 2000 to 2014, and use three existing parametric wind models, Holland10, Willoughby04, and CLE15, as the reference symmetric wind fields. The XGBoost-based wind reconstruction model improves the representation of both the surface wind asymmetries and the symmetric component. The interpretability of the model provides insights into the physical relationships of storm and environment features with hurricane wind asymmetries.